Algorithmic Fairness in Criminal Justice
Algorithms are used in the criminal justice system to streamline decisions in policing, pre-trial risk assessment and sentencing. Individuals are assigned risk scores that track their probability of committing future crimes or being crime victims. Risk scores are based on factors that correlate more or less strongly with crime, say prior arrests, prior convictions, failure to appear in court, involvement in crime incidents. To identify the relevant factors, different statistical and machine learning methods, such as logistic regression, are applied to historical data about crime. Validations methods ensure that algorithmic predictions about future crimes track actual outcomes.
Algorithms have taken the justice system by storm and are going to become even more pervasive. They promise to be efficient, objective and unbiased. But they are not without problems of course. Protected attributes, race or gender, are typically not part of the set of predictors, but if the training data are biased, algorithmic decisions will inherit the biases. Garbage in, garbage out. Biases in the data might compound over time through feedback loops and exacerbate racial inequities. Algorithms might actually fail at what they are supposed to do best, namely to reduce crime.
I will focus on the question of what it means for an algorithm to operate fairly. There is a growing literature in computer science, law and philosophy on algorithmic fairness. Definitions proliferate, but a unified analytical framework is lacking. I will discuss some of the proposals for theorizing about algorithmic fairness. Before I do that, I will review some examples of algorithms in criminal justice.
The Chicago Police Department, following a surge in street violence in 2013, adopted SSL (Strategic Subject List), an algorithm that calculates risk scores for individuals known to the police. The risk scores are intended to reflect someone's probability of being involved in a shooting incident either as a victim or an offender. The attributes used to calculate the risk scores are (see data in the Chicago Data Portal):
- times being the victim of a shooting incident
- age during the latest arrest
- times being the victim of aggravated battery or assault
- prior arrests for violent offenses
- gang affiliation
- prior narcotic arrests
- trend in involvement in crime incidents
- prior unlawful use of weapon arrests
Jessica Saunder and others in 2016 run an experimental analysis and concluded that “[i]ndividuals on the SSL are not more or less likely to become a victim of a homicide or shooting than the comparison group.”
SSL and its 2019 replacement SAID (Subject Assessment and Information Dashboard) rely on CVRM (Crime and Victimization Risk Model), a statistical model developed by the Illinois Institute of Technology that estimates an individual's risk of becoming a victim or offender. CVRM was already part of SSL, but its newer version in SAID relies on a shorter list of predictors, excluding previous narcotic arrests and gang activity. This is part of the larger Chicago Police Violence Reduction Strategy.
SSL and SAID are person-focused tools but there are also algorithms for predictive policing that are place-focused such as HunchLab, PredPol or CompStat. Based on geographical crime data and other factors, these tools help to allocate police resources toward certain areas rather than others.
Pre-trial risk assessment
Many algorithms are used to guide pre-trial risk assessment decisions. These decisions have to do with whether a defendant awaiting trial can be released or should be held in preventative custody to ensure public safety.
PSA, an algorithm developed by ArnoldVentures, is now widely used in several jurisdictions for this purpose. A report shows that the adoption of PSA in New Jersey as part of the Criminal Justice Reform has “reduced the length of time defendants spend in jail in the month following arrest.” There was a 20% reduction in the pre-trial jail population in 2017 after PSA was adopted. The ACLU of New Jersey endorsed the reform because it ended the pre-trial system of money and bail that disproportionately harms the poor.
But, perhaps, the most famous and discussed algorithmic tool for pre-trial crime prevention is COMPAS. It was at the center of a controversy about what it means for an algorithm to be fair.
The ProPublica/Northpointe debate
In 2016, ProPublica published a report asserting that COMPAS was racially biased against blacks. ProPublica compared the risk assessments made by COMPAS with the rearrest data for Broward County, Florida. Using rearrests as a proxy for recidivism, ProPublica found racial biases in COMPAS risk assessments.
COMPAS assigns to people a risk score between 1 and 10. For simplicity, think of COMPAS as a binary classifier, sorting people into two classes: those with a score of 6 or higher (considered high risk) and those with a score of 5 or lower (considered low risk). The analysis by ProPublica showed that blacks who were not rearrested were misclassified as high risk (a false positive) at almost twice the rate of similarly situated whites, 44.9% versus 23.5%. By contrast, blacks who were rearrested were misclassified as low risk (a false negative) at almost half the rate of whites, 28% versus 47.7%. These are damning numbers. COMPAS prefers to label blacks high risk even when they do not reoffend, and label whites low risk even when they do reoffend.
Northpointe (now Equivant), the company that designed COMPAS, responded that comparing false positives and false negatives across racial groups was irrelevant. What mattered were the predictions, and COMPASS made equally accurate predictions across racial groups. Among the individuals labeled high risk, the same proportion did not reoffend for both racial groups. Further, among the individuals labeled low risk, the same proportion did reoffend for both racial groups. COMPAS was not, Northpointe argued, racially biased.
The disagreement between Northpointe and ProPublica raises the question, what does algorithmic fairness consist in?
Definitions of algorithmic fairness
The computer science literature has identified several possible definitions – 21 in fact – by equalizing along different dimensions. Two conceptions of algorithmic fairness include:
- statistical parity, that is, the same proportion of members of different groups are coerced as a result of algorithmic decisions
- same threshold fairness, that is, the same risk threshold is applied to different groups
Statistical parity is sometimes rejected because if two groups have different rates of criminality, an accurate algorithm is expected to reflect this difference, and thus statistial parity will be violated.
Same threshold fairness is harder to reject. It is intuitively unfair to impose coercion on members of one group using a different standard compared to members of another group, say using a minimal score of 6 for blacks and a minimal score of 8 for whites. This might not be so obvious if different threhsolds reflect different levels of harm that coercion imposes upon different groups (more on this below).
Two other popular conceptions of algorithmic fairness emerge from the debate between ProPublica and Northpointe. ProPublica singled out the group of actual non-reoffenders (or reoffenders) and then compared the algorithm classification errors for whites and blacks within this group, for example, what percentage of white non-reoffenders are misclassified as high risk (a false positive) compared to the percentage of black non-reoffenders who are misclassified as high risk (a false positive again). Call equality along this dimension classification fairness or equality in false positives and false negatives.
Northpointe, instead, singled out the group of those labeled high risk (or low risk) by COMPAS and then compared the classification errors for whites and blacks within this group, for example, what percentage of white high risk individuals are non-reoffenders compared to the percentage of black high risk individuals who are non-reoffenders. Call equality along this dimension predictive fairness or equality in prediction errors.
The debate between ProPublica and Northpointe was about whether classification fairness or predictive fairenss should matter most. Both are plausible in their own way, but unfortunately there is no room for a conciliatory stance. As it turns out, it is a mathematical impossibility for an algorithm to ensure equality in false positive and false negative rates together with equality in prediction errors across groups. So long as the prevalence of recidivism or crime differs – as is the case between whites and blacks – the two dimensions of equality cannot be satisfied at the same time. We are confronted with an either/or choice.
Against classification fairness
Classification fairness is intuitively plausible. Take two individuals belonging to different racial groups, neither of whom is going to reoffend. They are equal relative to the outcome the algorithm should flag, namely recidivism. Why should one be more exposed to coercion than the other? Their membership in different groups should not put one at greater disadvantage than the other.
If one is a reoffender and the other is not, the algorhitm should treat the two differently. It should flag one as high risk but not the other. But classification fairness only requires that, regardless of group membership, those who are similarly situated – reoffeders or non-reoffenders – be exposed to coercion equally.
Sandra G. Mayson disagrees. In her 2019 article Bias In, Bias Out Yale Law Journal (128: 2218-2300), she argues that equality in false positives and false negatives does not matter for algorithmic fairness.
Her argument begins with two general premises:
(P1) Fairness requires to treat two similarly situated individuals the same.
This is Aristotle's principle of treating like cases alike. But when are two people “similarly situated”? Mayson writes:
The question of what makes two people (or groups) relevantly “alike” for purposes of a particular action is really a question about the permissible grounds for that action. To judge that two people with equivalent skill and experience are relevantly “alike” for purposes of a hiring decision is to judge that skill and experience are good grounds on which to make such a decision (2273).
The thought is that, in hiring decisions, two people with equal qualifications and work experience are similarly situated because work experience and qualifications are relevant grounds for making hiring decisions. That two people do not have the same ice-cream preferences does not matter as to whether they are similarly situated relative to hiring decisions because ice-cream preferences are not relevant for hiring decisions. The second premise is therefore:
(P2) Two individuals are similarly situated, relative to a certain decision, if and only if they share the features that are relevant grounds for making the decision in question.
Suppose we compare a white and a black individual who differ from one another in terms of the ultimate outcome, that is, one is a reoffender and the other is not. Does this difference count against them being similarly situated for the purpose of risk assessment? Mayson first notes that:
To hold that ultimate outcomes are what render two people (or groups) alike for purposes of risk assessment is to hold that outcomes are a good basis for risk assessment (2275).
That is: (P3) If outcomes such as being a reoffender (or being a non-reoffender) were to make two individuals similarly situated, then these outcomes would be relevant grounds for risk assessment.
(P3) follows from (P2) which posits a tight connection between features that are relevant decision-making grounds and features that make individuals similarly situated relative to the decision in question.
Mayson further adds that
they [=ultimate outcomes] cannot be the basis for risk assessment because at the time of assessment they are unknown. This is why we resort to risk assessment in the first place (2275).
So another premise of Mayson's argument reads:
(P4) Outcomes such as being a reoffender (or being a non-reoffender) are not relevant grounds for risk assessments.
From (P3) and (P4), it follows that ultimate outcomes cannot render two people similarly situated. Further, since by (P1) being similarly situated is relevant for fairness, ultimate outcomes are irrelevant for fairness. Finally, since the demand for equality in false positives and false negatives assumes that equality of outcomes renders two individuals similarly situated and thus is relevant for fairness, that demand is misplaced. In short,
The demand for equal algorithmic treatment for same-outcome groups amounts to a judgment that outcomes are the appropriate basis for prediction. And that judgment is nonsensical (2275).
This is an intriguing argument.
But I wonder whether Mayson is too narrowly focused on risk assessment. The latter is a tool for informing pre-trial decisions about detention or release. True, what's at stake is the fairness of algorithms, but more broadly, what's at stake is the fairness of decisions based on algorithms. It is odd to say that recidivism is not a legitimate ground for decisions about detention or release. It is actually the definitive ground. After all, if one will recidivate, they should be detained, and if not, they should be released.
Mayson might reply that, since outcomes are unknowable, they cannot be grounds for decision-making. It is unclear why that should be so. If two students are, in some objective sense, equally prepared, but a test systematically ranks one below the other, the test is unfair. Objective preparation might well be unknowable, but it would be the definitive ground that renders two students equally situated relative to test taking and assessment. Deborah Hellman in her forthcoming article Measuring Algorithmic Fairness, Virginia Law Review, uses this pedagogical example to argue that equality in false positives and false negatives matters for fairness. There are interesting parallels between fairness in standarized tests, a topic much discussed in the sixties, and today's debate about algorithmic fairness.
Against classification fairness (cont'ed)
Even if Mayson's argumet isn't right, there are other reasons to resist equality in false positives and false negatives as a criterion for fairness. Suppose false positives are higher for blacks than for whites, say, 40% v. 20%. One way to equalize them is to increase the risk threshold for imposing coercion on blacks, say from a risk score of 6 to 8, or conversely lower the threshold for whites, say from 6 to 4.
One immediate problem with this is that people in different groups will be classified high risk or low risk following different thresholds - a violation of same threshold fairness. And even if this is not regarded as problematic per se, there will also be another complication.
If a more stringent threshold is adopted for black defedants, more black reoffenders will be released. Since crime is often intraracial, this will adversely affect the black community almost exclusively. In the name of equality, the black community will be put at even greater disadvantage, other things equal (see, however, the argument below).
If a less stringent threshold is adopted for white defendants, nobody will be better off. Blacks will still be subject to a 40% false positive rate. The false positive rate for whites will increase from 20% to 40%. Whites will be worse off with no gain for blacks. This is the so-called leveling down objection to egalitarianism.
These problems can be avoided if false positives and false negatives can be equalized without adjusting the decision threshold. I believe this can be done, but I leave it for another time.
Equal protection jurisprudence
Suppose the preceeding arguments against classification fairness are correct. Are we to conclude that its alternative, predictive fairness, is the right notion? A larger problem remains. A clear analytical framework for algorithmic fairness is still missing. The computer science literature provides several possible definitions, but no clear reason why they matter for algorithmic fairness and to what extent. Defending one conception or another, in a piecemeal way, is unsatisfactory.
One obvious place to look for a general framework is equal protection jurisprudence, specifically, the notions of disparate treatment and disparate impact. Aziz Z. Huq has shown, however, this is not going to work; see his 2019 article Racial Equity in Algorithmic Criminal Justice, Duke Law Journal (68: 1043-1134).
Start with disparate treatment. Any differential treatment that is due to racial animus or discrminatory intent is clearly illegal. The use of race as a factor is not automatically prohibited, but must be justified with a clear rationale, an inquiry known as “strict scrutiny”. The Supreme Court has allowed the use of race in specified contexts, for example, in college admissions for the purpose of fostering diversity; see Fisher v. University of Texas (2016).
Disparate treatment is unlikely to be helpful here. Algorithms hardly have a racially discriminatory intent nor do they explicitly rely on race as a predictor. Perhaps, the concern is that using race sends a demeaning message and entrenches racial stereotypes. But the use of race in algorithms – if race is used at all – is usually opaque. No demeaning message can be clearly identified. Perhpas, the problem with racial classification is that using race in algorithms would fail to treat defendants as individuals. But so long as race is used with other factors or not used at all, this concern seems unjustified.
Huq makes a historical comment:
Current doctrinal approaches to constitutional racial equality … were configured in the context of judicial efforts to dismantle educational segregation in the Jim Crow South and then during a political backlash to the Civil Rights Movement … the legal conception of racial discrimination as a matter of intention or classification would reflect judicial concern with the discretionary choices of the police officer, school board president, or state legislator—that is, the modal problems presented by mid-century civil rights law (1101).
Today's context is much different:
A set of tools developed for a regulatory world of dispersed state actors, occasionally motivated by naked animus, cannot be mechanically translated into a world of centralized, computational decision-making (1103).
What about disparate impact as invoked in Title VII discrimination cases? Evidence of disparate impact against a protected group is enough to make a prima facie case of discrimination; see e.g. Hazelwood School District v. United States (1977). This applies to sectors such as employment and housing. The criminal justice system, however, seems exempt. In McClensky v. Kemp (1987), the US Supreme Court ruled that disparate racial impact is not enough to establish a constitutional violation. An elaborate statistical analysis – showing that death penalty decisions in Georgia disproportionately targeted African Americans, controlling for several variables – was not enough to convince the Court that the system violated equal protection.
Cost/benefit and multiple thresholds
If the legal framework is either obsolete or inapplicable, what else is there? Huq in his 2019 article favors a cost/benefit framework:
the key question for racial equity is whether the costs that an algorithmically driven policy imposes upon a minority group outweigh the benefits accruing to that group (1111).
A narrow view considers just immediate benefits (say, increased public safety) and immediate costs (say, unwarranted detention). But besides immediate costs, there are also more far reaching externalities:
[they] take many forms, including the effect of high incarceration rates on black communities and children as well as the social signification of race as a marker of criminality (1113)
These negative spillover effects on family life, employment and racial stigma, are likely to disproportionately affect minorities. As Huq writes:
the spillover costs of coercion of minority individuals for the minority group will be greater on a per capita basis than the costs of coercing majority group members (1113).
This observation has an important consequence:
accounting for both the immediate and spillover costs of crime control when its immediate benefits are small conduces to a bifurcated risk threshold—one rule for the majority, and one for minority (1131).
In other words, if decision thresholds should be set at a socially efficient level balancing the costs and benefits of pre-trial coercion, and if the costs of coercion are higher for blacks than for whites because of uneven spillover effects, then the standard for imposing coercion must be more stringent for blacks than for whites, other things equal.
This is a powerful argument. The emphasis on spillover costs has the merit of broadening the discussion about algorithmic fairness beyond a narrow focus on different equality metrics. Cost/benefit analysis offers a more principled way to think about the issue. But there are complications.
First, Huq emphasizes negative spillover effects and externalities, but he believes these are less relevant in the case of serious crimes:
… to focus solely on the immediate costs and benefits of a coercive intervention and to ignore externalities … seems a plausible approach with serious crimes, where externalities are dwarfed by immediate costs and benefits (1113).
When negative externalities can be ignored, the same threshold should presumably be imposed for whites and blacks for pre-trial coercion, other things being equal. But is this true of serious crimes? Serious crimes in some communities might have greater net spillover costs than in other communities. A wealthier family can more easily cope with the loss of an adult figure than a less wealthy family.
Huq's consequentialist framework would mandate that a less stringent threshold apply for algorithm-based coercion meant to prevent serious crimes in black communities so long as the prevention of serious crime in black communites has greater net benefits. On this view, coercing blacks at much higer rates than whites is justified provided coercion has positive net benefits for both communities.
The second complication is that the costs/benefit framework makes little difference between reoffenders and non-reoffenders. Huq's utilitarian calculus is not sensitive to this difference:
there is no particular reason to believe that any of these spillover costs are less if the person subject to the coercion is in fact a true rather than false positive … what should matter is the absolute cost of using a coercive tactic against a member of a minority group, net of benefit, for all members of that racial group. (1127-28)
A tacit notion of equality is at work here. It is unfair for a group, considered as a whole, to accrue positive benefits as a result of algorithmic coercion and for another group, again considered as a whole, to accrue negative net benefits. Call this expected utility equality. One wonders, why equalize along this dimension? By appealig to the cost/benefit framework, Huq has implicitly adopted yet another conception of algorithmic fairness.